models of ecological rationality: the recognition · pdf file · 2003-12-31models...

Models of Ecological Rationality: The Recognition Heuristic

Daniel G. Goldstein and Gerd GigerenzerMax Planck Institute for Human Development

One view of heuristics is that they are imperfect versions of optimal statistical procedures considered toocomplicated for ordinary minds to carry out. In contrast, the authors consider heuristics to be adaptivestrategies that evolved in tandem with fundamental psychological mechanisms. The recognition heuristic,arguably the most frugal of all heuristics, makes inferences from patterns of missing knowledge. Thisheuristic exploits a fundamental adaptation of many organisms: the vast, sensitive, and reliable capacityfor recognition. The authors specify the conditions under which the recognition heuristic is successful andwhen it leads to the counterintuitive less-is-more effect in which less knowledge is better than more formaking accurate inferences.

What are heuristics? The Gestalt psychologists Karl Dunckerand Wolfgang Koehler preserved the original Greek definition of“serving to find out or discover” when they used the term todescribe strategies such as “looking around” and “inspecting theproblem” (e.g., Duncker, 1935/1945). For Duncker, Koehler, anda handful of later thinkers, including Herbert Simon (e.g., 1955),heuristics are strategies that guide information search and modifyproblem representations to facilitate solutions. From its introduc-tion into English in the early 1800s up until about 1970, the termheuristics has been used to refer to useful and indispensablecognitive processes for solving problems that cannot be handled bylogic and probability theory (e.g., Polya, 1954; Groner, Groner, &Bischof, 1983).

In the past 30 years, however, the definition of heuristics haschanged almost to the point of inversion. In research on reasoning,judgment, and decision making, heuristics have come to denotestrategies that prevent one from finding out or discovering correctanswers to problems that are assumed to be in the domain ofprobability theory. In this view, heuristics are poor substitutes forcomputations that are too demanding for ordinary minds to carryout. Heuristics have even become associated with inevitable cog-nitive illusions and irrationality (e.g., Piattelli-Palmerini, 1994).

The new meaning of heuristics—poor surrogates for optimalprocedures rather than indispensable psychological tools—emerged in the 1960s when statistical procedures such as analysisof variance (ANOVA) and Bayesian methods became entrenchedas the psychologist’s tools. These and other statistical tools weretransformed into models of cognition, and soon thereafter cogni-tive processes became viewed as mere approximations of statisti-cal procedures (Gigerenzer, 1991, 2000). For instance, when WardEdwards (1968) and his colleagues concluded that human reason-ing did not accord with Bayes’s rule (a normative standard formaking probability judgments), they tentatively proposed that ac-tual reasoning is like a defective Bayesian computer with wrongly

combined values (misaggregation hypothesis) or misperceivedprobabilities (misperception hypothesis). The view of cognitiveprocesses as defective versions of standard statistical tools was notlimited to Edward’s otherwise excellent research program. In the1970s, the decade of the ANOVA model of causal attribution,Harold Kelley and his colleagues suggested that the mind at-tributes a cause to an effect in the same way that experimentersdraw causal inferences, namely, by computing an ANOVA:

The assumption is that the man in the street, the naive psychologist,uses a naive version of the method used in science. Undoubtedly, hisnaive version is a poor replica of the scientific one—incomplete,subject to bias, ready to proceed on incomplete evidence, and so on.(Kelley, 1973, p. 109)

The view that mental processes are “poor replicas” of scientifictools became widespread. ANOVA, multiple regression, first-order logic, and Bayes’s rule, among others, have been proposed asoptimal or rational strategies (see Birnbaum, 1983; Hammond,1996; Mellers, Schwartz, & Cooke, 1998), and the term heuristicswas adopted to account for discrepancies between these rationalstrategies and actual human thought processes. For instance, therepresentativeness heuristic (Kahneman & Tversky, 1996) wasproposed to explain why human inference is like Bayes’s rule withthe base rates left out (see Gigerenzer & Murray, 1987). Thecommon procedure underlying these attempts to model cognitiveprocesses is to start with a method that is considered optimal,eliminate some aspects, steps, or calculations, and propose that themind carries out this naive version.

We propose a different program of cognitive heuristics. Ratherthan starting with a normative process model, we start with fun-damental psychological mechanisms. The program is to design andtest computational models of heuristics that are (a) ecologicallyrational (i.e., they exploit structures of information in the environ-ment), (b) founded in evolved psychological capacities such asmemory and the perceptual system, (c) fast, frugal, and simpleenough to operate effectively when time, knowledge, and compu-tational might are limited, (d) precise enough to be modeledcomputationally, and (e) powerful enough to model both good andpoor reasoning. We introduce this program of fast and frugalheuristics here with perhaps the simplest of all heuristics: therecognition heuristic.

Correspondence concerning this article should be addressed to Daniel G.Goldstein and Gerd Gigerenzer, Center for Adaptive Behavior and Cog-nition, Max Planck Institute for Human Development, Lentzeallee 94,14195 Berlin, Germany. E-mail: [email protected] [email protected]

Psychological Review Copyright 2002 by the American Psychological Association, Inc.2002, Vol. 109, No. 1, 75–90 0033-295X/02/$5.00 DOI: 10.1037//0033-295X.109.1.75

75

In this article, we define the recognition heuristic and study itsbehavior by means of mathematical analysis, computer simulation,and experiment. We specify the conditions under which the rec-ognition heuristic leads to less-is-more effects: situations in whichless knowledge is better than more knowledge for making accurateinferences. To begin, we present two curious findings illustrate thecounterintuitive consequences of the recognition heuristic.

Can a Lack of Recognition Be Informative?

In the statistical analysis of experimental data, missing data arean annoyance. However, outside of experimental designs—whendata are obtained by natural sampling rather than systematic sam-pling (Gigerenzer & Hoffrage, 1995)—missing knowledge can beused to make intelligent inferences. We asked about a dozenAmericans and Germans, “Which city has a larger population: SanDiego or San Antonio?” Approximately two thirds of the Ameri-cans correctly responded that San Diego is larger. How manycorrect inferences did the more ignorant German group achieve?Despite a considerable lack of knowledge, 100% of the Germansanswered the question correctly. A similar surprising outcome wasobtained when 50 Turkish students and 54 British students madeforecasts for all 32 English F. A. Cup third round soccer matches(Ayton & Onkal, 1997). The Turkish participants had very littleknowledge about (or interest in) English soccer teams, whereas theBritish participants knew quite a bit. Nevertheless, the Turkishforecasters were nearly as accurate as the English ones (63% vs.66% correct).

At first blush, these results seem to be in error. How could moreknowledge be no better—or worse—than significantly less knowl-edge? A look at what the less knowledgeable groups knew mayhold the answer. All of the Germans tested had heard of SanDiego; however, about half of them did not recognize San Anto-nio. All made the inference that San Diego is larger. Similarly, theTurkish students recognized some of the English soccer teams (orthe cities that often make up part of English soccer team names)but not others. Among the pairs of soccer teams in which theyrated one team as completely unfamiliar and the other as familiarto some degree, they chose the more familiar team in 627 of 662cases (95%). In both these demonstrations, people used the factthat they did not recognize something as the basis for their pre-dictions, and it turned out to serve them well.

The strategies of the German and Turkish participants can bemodeled by what we call the recognition heuristic. The task towhich the heuristic is suited is selecting a subset of objects that isvalued highest on some criterion. An example would be to usecorporate name recognition for selecting a subset of stocks fromStandard and Poor’s 500, with profit as the criterion (Borges,Goldstein, Ortmann, & Gigerenzer, 1999). In this article, as in thetwo preceding laboratory experiments, we focus on the case ofselecting one object from two. This task is known as pairedcomparison or two-alternative forced choice; it is a stock-in-tradeof experimental psychology and an elementary case to which manyother tasks (such as multiple choice) are reducible.

The recognition heuristic is useful when there is a strong cor-relation—in either direction—between recognition and criterion.For simplicity, we assume that the correlation is positive. Fortwo-alternative choice tasks, the heuristic can be stated as follows:

Recognition heuristic: If one of two objects is recognized and theother is not, then infer that the recognized object has the higher valuewith respect to the criterion.

The recognition heuristic will not always apply, nor will it alwaysmake correct inferences. Note that the Americans and English inthe experiments reported could not apply the recognition heuris-tic—they know too much. It is also easy to think of instances inwhich an object may be recognized for having a small criterionvalue. Yet even in such cases the recognition heuristic still predictsthat a recognized object will be chosen over an unrecognizedobject. The recognition heuristic works exclusively in cases oflimited knowledge, that is, when only some objects—not all—arerecognized.

The effectiveness of the apparently simplistic recognition heu-ristic depends on its ecological rationality: its ability to exploit thestructure of the information in natural environments. The heuristicis successful when ignorance, specifically a lack of recognition, issystematically rather than randomly distributed, that is, when it isstrongly correlated with the criterion. (When this correlation isnegative, the heuristic leads to the inference that the unrecognizedobject has the higher criterion value.)

The direction of the correlation between recognition and thecriterion can be learned from experience, or it can be geneticallycoded. The latter seems to be the case with wild Norway rats.Galef (1987) exposed “observer” rats to neighbors that had re-cently eaten a novel diet. The observers learned to recognize thediet by smelling it on their neighbors’ breath. A week later, theobservers were fed this diet and another novel diet for the first timeand became ill. (They were injected with a nauseant, but as far asthey knew it could have been food poisoning.) When next pre-sented with the two diets, the observer rats avoided the diet thatthey did not recognize from their neighbors’ breath. Rats operateon the principle that other rats know what they are eating, and thishelps them avoid poisons. According to another experiment(Galef, McQuoid, & Whiskin, 1990), this recognition mechanismworks regardless of whether the neighbor rat is healthy or notwhen its breath is smelled. It may seem unusual that an animalwould eat the food its sick neighbor had eaten; however, rats seemto follow recognition without taking further information (such asthe health of the neighbor) into account. In this article, we inves-tigate whether people follow the recognition heuristic in a similarnoncompensatory way. If reasoning by recognition is a strategycommon to humans, rats, and other organisms, there should beaccommodations for it in the structure of memory. We explore thisquestion next.

The Capacity for Recognition

As we wander through a stream of sights, sounds, tastes, odors,and tactile impressions, some novel and some previously experi-enced, we have little trouble telling the two apart. The mechanismfor distinguishing between the novel and recognized seems to bespecialized and robust. For instance, recognition memory oftenremains when other types of memory become impaired. Elderlypeople suffering memory loss (Craik & McDowd, 1987; Schon-field & Robertson, 1966) and patients suffering certain kinds ofbrain damage (Schacter & Tulving, 1994; Squire, Knowlton, &Musen, 1993) have problems saying what they know about an

76 GOLDSTEIN AND GIGERENZER

object or even where they have encountered it. However, theyoften know (or can act in a way that proves) that they haveencountered the object before. Such is the case with R. F. R., a54-year-old policeman who developed such severe amnesia that hehad great difficulty identifying people he knew, even his wife andmother. One might be tempted to say he had lost his capacity forrecognition. Yet on a test in which he was shown pairs of photo-graphs consisting of one famous and one nonfamous person, hecould point to the famous persons as accurately as could a healthycontrol group (Warrington & McCarthy, 1988). His ability todistinguish between the unrecognized people (whom he had neverseen before) and the recognized people (famous people he hadseen in the media) remained intact. However, his ability to recallanything about the people he recognized was impaired. Laboratoryresearch has demonstrated that memory for mere recognition en-codes information even in divided-attention learning tasks that aretoo distracting to allow more substantial memories to be formed(Jacoby, Woloshyn, & Kelley, 1989). Because recognition contin-ues to operate even under adverse circumstances, and it can beimpaired independently from the rest of memory, we view it as aprimordial psychological mechanism (for cases involving a selec-tive loss of recognition, see Delbecq-Derousne, Beauvois, & Shal-lice, 1990).

Because the word recognition is used in different ways bydifferent researchers, it needs to be precisely defined. Considerthree levels of knowledge an organism may have. First, one mayhave no knowledge of an object or event because one has neverheard, smelled, touched, tasted, or seen it before. Such objects wecall “unrecognized.” Second are objects one has experienced be-fore but of which one has absolutely no further knowledge beyondthis initial sense of recognition. These we will call “merely rec-ognized.” (The Scottish verb “to tartle” gives a name to this stateof memory; people tartle when they recognize another’s face butcannot remember anything else about him or her.) The third levelof knowledge comprises mere recognition plus further knowledge;not only does one recognize the object, but one can provide allsorts of additional information about it, such as where one encoun-tered it, what it is called, and so on. Thus, with the term recogni-tion, we divide the world into the novel and the previouslyexperienced.

This use of the term needs to be distinguished from another use,which might be characterized as “recognition of items familiarfrom a list” (e.g., Brown, Lewis, & Monk, 1977). Here the behav-ior of interest is a person’s ability to verify whether a commonthing (usually a word such as house) had been presented in aprevious experimental session. Studies dealing with this meaningof recognition often fail to touch on the distinction between thenovel and the previously experienced because the stimuli—mostlynumbers or everyday words—are certainly not novel to the par-ticipant before the experiment. Experiments that use nonwords ornever-seen-before photographs capture the distinction of interesthere: that between the truly novel and the previously experienced.

Recognition also needs to be distinguished from notions such asavailability (Tversky & Kahneman, 1974) and familiarity (Griggs& Cox, 1982). The availability heuristic is based on recall, notrecognition. People recognize far more items than they can recall.Availability is a graded distinction among items in memory and ismeasured by the order or speed with which they come to mind orthe number of instances of categories one can generate. Unlike

availability, the recognition heuristic does not address comparisonsbetween items in memory, but rather the difference between itemsin and out of memory (Goldstein, 1997). The term familiarity istypically used in the literature to denote the degree of knowledge(or amount of experience) a person has of a task or object. Therecognition heuristic, in contrast, treats recognition as a binary,all-or-none distinction; further knowledge is irrelevant.

A number of studies demonstrate that recognition memory isvast, easily etched on, and remarkably retentive despite shortpresentation times. Shepard’s (1967b) experiment with 612 pairsof novel photographs shown with unlimited presentation timeresulted in participants correctly recognizing them with 99% ac-curacy. This impressive result was made to appear ordinary byStanding’s experiments 6 years later. Standing (1973) increasedthe number of pictures (photographs and “striking” photographsselected for their vividness) to 1,000 and limited the time ofpresentation to 5 s. Two days later, he tested recognition memorywith pairs of pictures in which one had been presented previouslyand one was novel. Participants were able to point to the previ-ously presented picture 885 times of 1,000 with normal picturesand 940 times of 1,000 with “striking” pictures. Standing thenoutdid himself by a factor of 10. In perhaps the most extensiverecognition memory test ever performed, he presented 10,000normal pictures for 5 s each. Two days later participants correctlyidentified them in 8,300 of 10,000 pairs. When more and morepictures were presented, the retention percentage declined slightly,but the absolute number of pictures recognized increased.

Standing’s participants must have felt they were making a fairnumber of guesses. Some research suggests that people may notrely on recognition-based inferences in situations in which theyfeel their memory is not reliable, such as when presentation timesare short or when there are distractions at presentation (Strack &Bless, 1994). However, Standing’s results, when adjusted by theusual guessing correction, come out well above chance. Withrespect to the performance with his “striking” pictures, Standingspeculated, “If one million items could be presented under theseconditions then 731,400 would be retained” (p. 210). Of course,presenting 1 million items is a feat no experimental psychologist islikely to try.

Recognition memory is expansive, sensitive, and reliableenough to serve in a multitude of inferential tasks. We now discusshow the accuracy of these recognition-based inferences relies notonly on the soundness of memory but also on the relationshipbetween recognition and the environment.

Theory

Recognition and the Structure of the Environment

In this article, we investigate recognition as it concerns propernames. Proper name recognition is of particular interest because itconstitutes a specialized domain in the cognitive system that canbe impaired independently of other language skills (McKenna &Warrington, 1980; Semenza & Sgaramella, 1993; Semenza &Zettin, 1989). Because an individual’s knowledge of geographycomprises an incomplete set of proper names, it is ideal forrecognition studies. In this article, we focus on two situations oflimited knowledge: Americans’ recognition of German city namesand Germans’ recognition of city names in the United States. TheAmerican students we have tested over the years recognized about

77ECOLOGICAL RATIONALITY: THE RECOGNITION HEURISTIC

one fifth of the 100 largest German cities, and the German studentsrecognized about one half of the 100 largest U.S. cities. Anotherreason cities were used to study proper name recognition is be-cause of the strong correlation between city name recognition andpopulation.

It should be clear, however, that the recognition heuristic doesnot apply everywhere. The recognition heuristic is domain specificin that it works only in environments in which recognition iscorrelated with the criterion being predicted. Yet, in cases ofinference, the criterion is not immediately accessible to the organ-ism. How can correlations between recognition and inaccessiblecriteria arise? Figure 1 illustrates the ecological rationality of therecognition heuristic. There are “mediators” in the environmentthat have the dual property of reflecting (but not revealing) thecriterion and being accessible to the senses. For example, a personmay have no direct information about the endowments of univer-sities, because this information is often confidential. However, theendowment of a university may be reflected in how often it ismentioned in the newspaper. The more often a name occurs in thenewspaper, the more likely it is that a person will hear of thisname. Because the newspaper serves as a mediator, a person canmake an inference about the inaccessible endowment criterion.Three variables reflect the strength of association between thecriterion, mediator, and recognition memory: the ecological cor-relation, the surrogate correlation, and the recognition validity.

The correlation between the criterion and the mediator is calledthe ecological correlation. In our example, the criterion is theendowment of a university and the mediator variable is the numberof times the university is mentioned in the newspaper. The surro-gate correlation is that between the mediator (a surrogate for theinaccessible criterion) and the contents of recognition memory, forinstance, the number of mentions in the newspaper correlatedagainst recognition of the names mentioned. Surrogate correlationscan be measured against the recognition memory of one person (inwhich case the recognition data will be binary) or against thecollective recognition of a group, which we will examine later.

Figure 1 is reminiscent of Brunswik’s lens model (e.g., Hammond,1996). Recall that the lens model has two parts: environmental andjudgmental. Figure 1 can be seen as an elaboration of the envi-ronmental part, in which recognition is a proximal cue for thecriterion and the recognition validity corresponds to Brunswik’s“ecological validity.” In contrast to the standard lens model, thepathway from the criterion to recognition is modeled by theintroduction of a mediator.

The single most important factor for determining the usefulnessof the recognition heuristic is the strength of the relationshipbetween the contents of recognition memory and the criterion. Wedefine the recognition validity as the proportion of times a recog-nized object has a higher criterion value than an unrecognizedobject in a reference class. The recognition validity � is thus:

� � R/(R � W),

where R is the number of correct (right) inferences the recognitionheuristic would achieve, computed across all pairs in which oneobject is recognized and the other is not, and W is the number ofincorrect (wrong) inferences under the same circumstances. Therecognition validity is essential for computing the accuracy attain-able through the recognition heuristic.

Accuracy of the Recognition Heuristic

How accurate is the recognition heuristic? What is the propor-tion of correct answers one can expect to achieve using therecognition heuristic on two-alternative choice tasks? Let us posita reference class of N objects and a test whose questions are pairsof objects drawn randomly from the class. Each of the objects iseither recognized by the test taker or unrecognized. The test scoreis the proportion of questions in which the test taker correctlyidentifies the larger of the two objects.

Consider that a pair of objects must be one of three types: bothrecognized, both unrecognized, or one recognized and one not.Supposing there are n recognized objects, there are N � n unrec-ognized objects. This means there are n(N � n) possible pairs inwhich one object is recognized and the other is not. Similarly,there are (N � n)(N � n � 1)/2 pairs in which neither object isrecognized. Both objects are recognized in n(n � 1)/2 pairs.Dividing each of these terms by the total number of possible pairs,N(N � 1)/2, gives the proportion of each type of question in anexhaustive test of all possible pairs.

The proportion correct on an exhaustive test is calculated bymultiplying the proportion of occurrence of each type of questionby the probability of scoring a correct answer on questions of thattype. The recognition validity �, it may be recalled, is the proba-bility of scoring a correct answer when one object is recognizedand the other is not. When neither object is recognized, a guessmust be made, and the probability of a correct answer is 1/2.Finally, � is the knowledge validity, the probability of getting acorrect answer when both objects are recognized. Combiningterms in an exhaustive pairing of objects, the expected proportionof correct inferences, f(n), is

f(n) �

2� n

N��N � n

N � 1�� N � n

N ��N � n � 1

N � 1 � 1

2� �n

N��n � 1

N � 1��.

(1)

Figure 1. The ecological rationality of the recognition heuristic. Aninaccessible criterion (e.g., the endowment of an institution) is reflected bya mediator variable (e.g., the number of times the institution is mentionedin the news), and the mediator influences the probability of recognition.The mind, in turn, uses recognition to infer the criterion.


Consider the three parts of the right side of the equation. Theterm on the left accounts for the correct inferences made by therecognition heuristic. The term in the middle represents the correctinferences resulting from guessing. The right-most term equals theproportion of correct inferences made when knowledge beyondrecognition is used. Note that if n � 0, that is, no objects arerecognized, then all questions will lead to guesses, and the pro-portion correct will be 1/2. If all objects are recognized (n � N),then the left-most two terms become zero and the proportioncorrect becomes �. The left-most term shows that the recognitionheuristic comes into play most under “half ignorance,” that is,when the number of recognized objects n equals the number ofunrecognized objects N � n. (Note, however, that this does notimply that proportion correct will be maximized under these con-ditions.) Equation 1 specifies the proportion of correct inferencesmade by using the recognition heuristic whenever possible basedon the recognition validity �, the knowledge validity �, and thedegree of recognition (n compared with N). Next, we shall see howit leads to a curious state in which less recognition is better thanmore for making accurate inferences.

The Less-Is-More Effect

Equation 1 seems rather straightforward but has some counter-intuitive implications. Consider the following thought experiment:Three Parisian sisters receive the bad news that they have to takea test on the 100 largest German cities at their lycee. The test willconsist of randomly drawn pairs of cities, and the task will be tochoose the more populous city. The youngest sister does not getout much; she has never heard of Germany (nor any of its cities)before. The middle sister is interested in the conversations ofgrown-ups and often listens in at her parents’ cocktail parties. Sherecognizes half of the 100 largest cities from what she has over-heard in the family salon. The elder sister has been furiouslystudying for her baccalaureate and has heard of all of the 100largest cities in Germany. The city names the middle sister hasoverheard belong to rather large cities. In fact, the 50 cities sherecognizes are larger than the 50 cities she does not recognize inabout 80% of all possible pairs (the recognition validity � is .8).The middle and elder sisters not only recognize the names of citiesbut also have some knowledge beyond recognition. When theyrecognize two cities in a pair, they have a 60% probability ofcorrectly choosing the larger one; that is, the knowledge validity �is .6, whereas � � .5 would mean they have no useful furtherknowledge. Which of the three sisters will score highest on the testif they all rely on the recognition heuristic whenever they can?Equation 1 predicts their performance as shown in Figure 2.

The youngest sister can do nothing but guess on every question.The oldest sister relies on her knowledge (�) on every question andscores 60% correct. How well does the middle sister, who has halfthe knowledge of her older sister, do? Surprisingly, she scores thegreatest proportion of correct inferences (nearly 68% correct,according to Equation 1). Why? She is the only one who can usethe recognition heuristic. Furthermore, she can make the most ofher ignorance because she happens to recognize half of the cities,and this allows her to use the recognition heuristic most often. Therecognition heuristic leads to a paradoxical situation in whichthose who know more exhibit lower inferential accuracy than those

who know less. We call such situations less-is-more effects. Figure 2shows how less-is-more effects persist with other values of �.

Forecasting Less-Is-More Effects

In what situations will less-is-more effects arise? We first spec-ify a sufficient condition in idealized environments and then usecomputer simulation in a real-world environment.

Equation 1 models the accuracy resulting from using the recog-nition heuristic. The less-is-more effect can be defined as the stateof affairs in which the value of f(n) in Equation 1 at some integerfrom 0 to N � 1 (inclusive) is greater than the value at N. For thisto happen, the maximum value of �(n), the continuous parabolaconnecting the discrete points, must occur closer to one of thepoints 0 to N � 1 than to point N, that is, between 0 and N � 1/2.Solving the equation ��(n) � 0, when ��(n) is simply the firstderivative of �(n), one locates the maximum of �(n) at

� �1 � 2� � 2N � 4�N)

2(1 � 4� � 2�). (2)

A simple calculation shows that when � � �, the location of themaximum of �(n) is equal to N � 1/2: exactly between the N � 1stand Nth points. Either increasing � or decreasing � from this pointcauses the fraction (Equation 2), and thus the location of themaximum, to decrease. From this, we can conclude that there willbe a less-is-more effect whenever � � �, that is, whenever theaccuracy of mere recognition is greater than the accuracy achiev-able when both objects are recognized.

This result allows us to make a general claim. Under theassumption that � and � are constant, any strategy for solving

Figure 2. Less-is-more effects illustrated for a recognition validity � �.8. When the knowledge validity � is .5, .6, or .7, a less-is-more effectoccurs. A � of .5 means that there is no knowledge beyond recognition.When the knowledge validity equals the recognition validity (.8), noless-is-more effect is observed; that is, performance increased with increas-ing n. The performance of the three sisters is indicated by the three pointson the curve for � � .6. The curve for � � .6 has its maximum slightly tothe right of the middle sister’s score.


multiple-choice problems that follows the recognition heuristicwill yield a less-is-more effect if the probability � of getting acorrect answer merely on the basis of recognition is greater thanthe probability � of getting a correct answer using moreinformation.

In this derivation, we have supposed that the recognition validity� and knowledge validity � remain constant as the number ofcities recognized, n, varies. Figure 2 shows many individuals withdifferent knowledge states but with fixed � and �. In the realworld, the recognition and knowledge validities usually vary whenone individual learns to recognize more and more objects fromexperience.1 Will a less-is-more effect arise in a learning situationin which � and � vary with n, or is the effect limited to situationsin which these are constant?

The Less-Is-More Effect: A Computer Simulation

To test whether a less-is-more effect might arise over a naturalcourse of learning, we ran a computer simulation that learned thenames of German cities in the order a typical American mightcome to recognize them. How can one estimate this order? Wemade the simplifying assumption that the most well-known citywould probably be learned about first, then the second-most well-known city, and so on. We judged how well-known cities are bysurveying 67 University of Chicago students and asking them toselect the cities they recognized from a list. These cities were thenranked by the number of people who recognized them.2 Munichturned out to be the most well-known city, so our computerprogram learned to recognize it first. Recognizing only Munich,the program was then given an exhaustive test consisting of allpairs of cities with more than 100,000 inhabitants (the same 83cities used in Gigerenzer & Goldstein, 1996). Next, it learned torecognize Berlin (the second-most well-known city) to make atotal of two recognized cities and was tested on all possible pairsagain. It learned and was tested on city after city until it recognizedall of them. In one condition, the program learned only the namesof cities and made all inferences by the recognition heuristic alone.This result is shown as the bottom curve in Figure 3 labeled NoCues. When all objects were unrecognized or all were recognized,performance was at a chance level. Over the course of learning, aninverse U shape, like that in Figure 2, reappears. Here the less-is-more curve is jagged because, as mentioned, the recognition va-lidity was not set to be a constant but was allowed to vary freelyas more cities were recognized.

Would the less-is-more effect disappear if the program learnednot just the names of cities but information useful for predictingcity populations as well? In other words, if there is recall ofrelevant facts, will this override the less-is-more effect and, there-fore, the recognition heuristic? In a series of conditions, the pro-gram learned the name of each city, along with one, two, or ninepredictive cues for inferring population (the same cues as inGigerenzer & Goldstein, 1996). In the condition with one cue, theprogram learned whether each of the cities it recognized was oncean exposition site, a predictor with a high ecological validity (.91;see Gigerenzer & Goldstein, 1996). The program then used adecision strategy called Take The Best (Gigerenzer & Goldstein,1999) to make inferences about which city is larger. Take The Bestis a fast and frugal strategy for drawing inferences from cues thatuses the recognition heuristic as the first step. It looks up cues in

the order of their validity and stops search as soon as positiveevidence for one object (e.g., the city was an exposition site) butnot for the other object is found. It is about as accurate as multipleregression for this task (Gigerenzer & Goldstein, 1999).

Does adding predictive information about exposition sites washout the less-is-more effect? With one cue, the peak of the curveshifted slightly to the right, but the basic shape persisted. When theprogram recognized more than 58 cities, including informationabout exposition sites, the accuracy still went down. In the condi-tion with two cues, the program learned the exposition site infor-mation and whether each city had a soccer team in the majorleague (another cue with a high validity, .87). The less-is-moreeffect was lessened, as is to be expected when adding recallknowledge to recognition, but still pronounced. Recognizing all

1 When there are many different individuals who recognize differentnumbers of objects, as with the three sisters, it is possible that eachindividual has roughly the same recognition validity. For instance, for theUniversity of Chicago students we surveyed (Gigerenzer & Goldstein,1996), recognition validity in the domain of German cities was about 80%regardless of how many cities each individual recognized. However, whenone individual comes to recognize more and more objects, the recognitionvalidity can change with each new object recognized. Assume a personrecognizes n of N objects. When she learns to recognize object n � 1, thiswill change the number of pairs for which one object is recognized and theother is not from n(N � n) to (n � 1) (N � n � 1). This number is thedenominator of the recognition validity (the number R � W of correct plusincorrect inferences), which consequently changes the recognition validityitself. Learning to recognize new objects also changes the numerator (thenumber R of correct inferences); in general, if the new object is large, therecognition validity will go down, and if it is small, it will go up.

2 If several cities tied on this measure, they were ordered randomly foreach run of the simulation.

Figure 3. Less-is-more effects when the recognition validity is not heldconstant but varies empirically, that is, as cities become recognized in orderof how well known they are. Inferences are made on recogniton alone (nocues) or with the aid of 1, 2, or 9 predictive cues and Take The Best (seeGoldstein & Gigerenzer, 1999).


cities and knowing all the information contained in two cues (thefar right-hand point) resulted in fewer correct inferences thanrecognizing only 23 cities. Finally, we tested a condition in whichall nine cues were available to the program, more information forpredicting German city populations than perhaps any Germancitizen knows. We see the less-is-more effect finally flattening out;however, it does not go away completely. Even when all 83 cuevalues are known for each of nine cues and all cities are recog-nized, the point on the far right is still lower than 24 other pointson that curve. A beneficial amount of ignorance can enable evenhigher accuracy than extensive knowledge can.

To summarize, the simplifying assumption that the recognitionvalidity � and knowledge validity � remain constant is not nec-essary for the less-is-more effect to arise: It held as � and � variedin a realistic way. Moreover, the counterintuitive effect evenappeared when complete knowledge about several predictors waspresent. The effect appears rather robust in theory and simulation.Will the recognition heuristic and, thus, the less-is-more effectemerge in human behavior?

Does the Recognition Heuristic PredictPeople’s Inferences?

It could be that evolution has overlooked the inferential ease andaccuracy the recognition heuristic affords. Do human judgmentsactually accord with the recognition heuristic? We test this ques-tion in an experiment in which people make inferences fromlimited knowledge.

Method

The participants were 22 students from the University of Chicago whowere native speakers of English and who had lived in the United States forthe last 10 years. They were given all the pairs of cities drawn from the 25(n � 6) or 30 (n � 16) largest cities in Germany, which resulted in 300 or435 questions for each participant, respectively. The task was to choose thelarger city in each pair. Furthermore, each participant was asked to checkthe names of the cities he or she recognized off of a list. Half of theparticipants took this recognition test before the experiment and half after.(Order turned out to have no effect.) From this recognition information, wecalculated how often each participant had an opportunity to choose inaccordance with the recognition heuristic and compared this number withhow often they actually did. If people use the recognition heuristic, theyshould predominantly choose recognized cities over unrecognized ones. Ifthey use a compensatory strategy that takes more information into account,this additional information may often suggest not choosing the recognizedcity. After all, a city can be recognized for reasons that have nothing to dowith population.

Results

Figure 4 shows the results for the 22 individual participants. Foreach participant, one bar is shown. The bar shows the proportionof judgments that agreed with the recognition heuristic among allcases in which it could be applied. For example, participant A had156 opportunities to choose according to the recognition heuristicand did so every time, participant B did so 216 of 221 times, andso on. The proportions of recognition heuristic adherence rangedbetween 73% and 100%. The mean proportion of inferences inaccordance with the recognition heuristic was 90% (median 93%).

This simple test of the recognition heuristic showed that it capturedthe vast majority of inferences.

Test Size Influences Performance

Equation 1 allows for a novel prediction: The number of correctinferences depends in a nonmonotonic but systematic way on thenumber of cities N included in the test. That is, for constantrecognition validity � and knowledge validity � the test size N(and n that depends on it) predicts various proportions of correctanswers. Specifically, Equation 1 predicts when the deletion oraddition of one object to the test set should decrease or increaseperformance.

Method

Using Equation 1, we modeled how the accuracy of the participants inthe preceding experiment would change if they were tested on variousnumbers of cities and test these predictions against the participants’ dem-onstrated accuracy. These predictions were made using nothing but infor-mation about which cities people recognize; no parameters are fit. With thedata from the previous experiment, we looked at the participants’ accuracywhen tested on the 30 largest cities, then on just the 29 largest cities, andso on, down to the 2 largest. (This was done by successively eliminatingquestions and rescoring.) For each participant, at each test size, we couldcompute the number of objects they recognized and the recognition validity� from the recognition test. Assuming, for simplicity, that the knowledgevalidity � always was a dummy value of .5, we used Equation 1 to predictthe change in the proportion of correct inferences when the number ofcities N on the test varied.

Results

The average predictions for and the average actual performanceof the individuals who took the exhaustive test on the 30 largestcities are shown in Figure 5. There are 28 predicted changes (up ordown), and 26 of these 28 predictions match the data. Despite the

Figure 4. Recognition heuristic accordance. For each of 22 participants,the bars show the percentage of inferences consistent with the recognitionheuristic. The individuals are ordered from left to right according to howoften their judgments agreed with the recognition heuristic.


apparent irregularity of the actual changes, Equation 1 predictsthem with great precision. Note that there are no free parametersused to fit the empirical curve.

An interesting feature of Figure 5 is the vertical gap between thetwo curves. This difference reflects the impact of the knowledgevalidity �, which was set at the dummy value of .5 for demon-strative purposes. The reason this difference decreases with in-creasing N is that, as the tests begin to include smaller and smallercities, the true knowledge validity � of our American participantstended toward .5. That is, it is safe to assume that when Americansare tested on the 30 (or more) largest German cities, their proba-bility of correctly inferring which of two recognized cities is largeris only somewhat better than chance.

To summarize, the recognition heuristic predicts how perfor-mance changes with increasing or decreasing test size. Theempirical data showed a strong, nonmonotonic influence onperformance, explained almost entirely by the recognitionheuristic.

Noncompensatory Inferences: Will Inference Follow theRecognition Heuristic Despite Conflicting Evidence?

People often look up only one or two relevant cues, avoidsearching for conflicting evidence, and use noncompensatory strat-egies (e.g., Einhorn, 1970; Einhorn & Hogarth, 1981, p. 71;Fishburn, 1974; Hogarth, 1987; Payne, Beltman, & Johnson, 1993;Shepard, 1967a). The recognition heuristic is a noncompensatorystrategy: If one object is recognized and the other is not, then theinference is determined; no other information about the recognizedobject is searched for and, therefore, no other information canreverse the choice determined by recognition. Noncompensatoryjudgments are a challenge to traditional ideals of rationality be-cause they dispense with the idea of compensation by integration.For instance, when Keeney and Raiffa (1993) discussed lexico-graphic strategies—the prototype of noncompensatory rules—they

repeatedly inserted warnings that such a strategy “is more widelyadopted in practice than it deserves to be” because “it is naivelysimple” and “will rarely pass a test of ‘reasonableness’”(pp. 77–78). The term lexicographic means that criteria are looked up in afixed order of validity, like the alphabetic order used to arrangewords in a dictionary. Another example of a lexicographic struc-ture is the Arabic (base 10) numeral system. To decide which oftwo numbers (with an equal number of digits) is larger, one looksat the first digit. If the first digit of one number is larger, then thewhole number is larger. If the first digits are equal, then one looksat the second digit, and so on. This simple method is not possiblefor Roman numerals, which are not lexicographic. Lexicographicstrategies are noncompensatory because the decision made on thebasis of a cue higher up in the order cannot be reversed by the cueslower in the order. The recognition heuristic is possibly the sim-plest of all noncompensatory strategies: It only relies on subjectiverecognition and not on objective cues. In this section, we are notconcerned with the “reasonableness” of its noncompensatory na-ture (which we have analyzed in earlier sections in terms of �, �,n, and N), but with the descriptive validity of this property. Wouldthe people following the recognition heuristic still follow it if theywere taught information that they could use to contradict thechoice dictated by recognition?

In this experiment, we used the same task as before (inferringwhich of two cities has the larger population) but taught participantsadditional, useful information that offered an alternative to followingthe recognition heuristic, in particular, knowledge about which citieshave soccer teams in the major league (the German “Bundesliga”).German cities with such teams tend to be quite large, so the presenceof a major league soccer team indicates a large population. Because ofthis relationship, we can test the challenging postulate of whether therecognition heuristic is used in a noncompensatory way. Which wouldparticipants choose as larger: an unrecognized city or a recognizedcity that they learned has no soccer team?

Figure 5. The use of the recognition heuristic implies that accuracy depends on test size (N) in an irregular butpredictable way (Equation 1). The predictions were made with recognition information alone, and no parameterswere fit. The knowledge validity used was a dummy value that assumes people will guess when both cities arerecognized. The predictions mirror the fluctuations in accuracy in 26 of the 28 cases.


Method

Participants were 21 students from the University of Chicago. All werenative English speakers who had lived in the United States for at least thelast 10 years. The experiment consisted of a training session and a testsession. At the beginning of the training session, participants were in-structed to write down everything they were taught, and they were in-formed that after training they would be given a test consisting of pairs ofcities drawn from the 30 largest in Germany. During the training session,participants were taught that 9 of the 30 largest cities in Germany havesoccer teams and that the 9 cities with teams are larger than the 21 citieswithout teams in 78% of all possible pairs. They were also taught thenames of 4 well-known cities that have soccer teams as well as the namesof 4 well-known cities that do not. When they learned about these 8 cities,they believed they had drawn them at random from all 30 cities; inactuality, the computer program administering the experiment was riggedto present the same information to all participants. After the training,participants were asked to recall everything they had been taught withouterror. Those who could not do so had to repeat training until they could.

After participants passed the training phase, they were presented pairs ofcities and asked to choose the larger city in each pair. Throughout this test,they could refer to their notes about which cities do and do not have soccerteams. To motivate them to take the task seriously, they were offered achance of winning $15 if they scored more than 80% correct. To reiterate,the point of the experiment was to see which city the participants wouldchoose as the larger one: a city they had never heard of before, or one thatthey had recognized beforehand but had just learned had no soccer team.From the information presented in the training session (which did not makeany mention of recognition), one would expect the participants to choosethe unrecognized city in these cases. The reason for this is as follows. Anunrecognized city either does or does not have a soccer team. If it does (a 5in 22 chance from the information presented), then there is a 78% chancethat it is larger. If it does not, then soccer team information is useless anda guess must be made. Any chance of the unrecognized city having a soccerteam suggests that it is probably larger. Participants who do not put anyvalue on recognition should always choose the unrecognized city.

Note that the role of recognition or the recognition heuristic was nevermentioned in the experiment. All instruction concerned soccer teams. Thedemand characteristics in this experiment would suggest that, after passingthe training session requirements, the participants would use the soccerteam information for making the inferences.

Results

The test consisted of 66 pairs of cities. Of these, we were onlyinterested in 16 critical pairs that contained one unrecognized cityand one recognized city that does not have a soccer team. Beforeor after this task, we tested which cities each participant recog-nized (order had no effect). In those cases in which our assump-tions about which cities the participants recognized were contra-dicted by the recognition test, items were eliminated from theanalysis, resulting in fewer than 16 critical pairs.3

Figure 6 reads the same as Figure 4. Twelve of 21 participantsmade choices in accordance with the recognition heuristic withoutexception, and most others deviated on only one or two items. Allin all, inferences accorded with the recognition heuristic in 273 ofthe 296 total critical pairs. Despite the presence of conflictingknowledge, the mean proportion of inferences agreeing with theheuristic was 92% (median 100%). These numbers are even a bithigher than in the previous study, which, interestingly, did notinvolve the teaching of contradictory information.

It appears that the additional information about soccer teamswas not integrated into the inferences, consistent with the recog-

nition heuristic. This result supports the hypothesis that the rec-ognition heuristic was applied in a noncompensatory way.

Will a Less-Is-More Effect Occur Between Domains?

A less-is-more effect can emerge in at least three differentsituations. First, it can occur between two groups of people, whena more knowledgeable group makes systematically worse infer-ences than a less knowledgeable group in a given domain. Anexample was the performance of the American and German stu-dents on the question of whether San Diego or San Antonio islarger. Second, a less-is-more effect can occur between domains,that is, when the same group of people achieves higher accuracy ina domain in which they know little than in a domain in which theyknow a lot. Third, a less-is-more effect can occur during knowl-edge acquisition, that is, when the same group or individual makesmore erroneous inferences as a result of learning. In this and thefollowing experiment, we attempt to demonstrate less-is-moreeffects of the latter two types, starting with a less-is-more effectbetween domains.

The mathematical and simulation results presented previouslyshow that less-is-more effects emerge under the conditions spec-ified. However, the curious phenomenon of a less-is-more effect isharder to demonstrate with real people than by mathematical proofor computer simulation. The reason is that real people do notalways need to make inferences under uncertainty; they sometimes

3 Another precaution we took concerned the fact that unrecognized citiesare often smaller than recognized ones, and this could work to the advan-tage of the recognition heuristic in this experiment. How can one tellwhether people are following the recognition heuristic or choosing cor-rectly by some other means? To prevent this confusion, the critical testitems were designed so that the unrecognized cities were larger than therecognized cities in half of the pairs.

Figure 6. Recognition heuristic adherence despite training to encouragethe use of information other than recognition. The bars show the percentageof inferences consistent with the recognition heuristic. The individuals areordered from left to right by recognition heuristic accordance.


have definite knowledge and can make deductions (e.g., if theyknow for certain that New York is the largest American city, theywill conclude that every other city is smaller). For this reason, evenif there is a between-domains less-is-more effect for all itemsabout which an inference must be made, this effect may be hiddenby the presence of additional, definite knowledge.

In the following experiment, American participants were testedon their ability to infer the same criterion, population, in twodifferent domains: German cities and American cities. Naturally,we expected the Americans to have considerably more knowledgeabout their own country than about Germany. Common sense (andall theories of knowledge of which we are aware) predicts thatparticipants will make more correct inferences in the domain aboutwhich they know more. The recognition heuristic, however, couldpull performance in the opposite direction, although its effect willbe counteracted by the presence of certain knowledge that theAmericans have about cities in the United States. Could the testscores on the foreign cities nevertheless be nearly as high as thoseon the domestic ones?

Method

Fifty-two University of Chicago students took two tests each: one onthe 22 largest cities in the United States, and one on the 22 largest cities inGermany. The participants were native English speakers who had lived thepreceding 10 years in the United States. Each test consisted of 100 pairs ofrandomly drawn cities, and the task was to infer which city is the larger ineach pair. Half the subjects were tested on the American cities first and halfon the German cities first. (Order had no effect.)

As mentioned, the curious phenomenon of a less-is-more effect is harderto demonstrate with real people than on paper because of definite knowl-edge. For instance, many Americans, and nearly all of the University ofChicago students, can name the three largest American cities in order.Knowing only the top three cities and guessing on questions that do notinvolve them led to 63% correct answers without making any inferences,only deductions. Knowing the top five in order yields 71% correct. Nocomparable knowledge can be expected for German cities. (Many of ourparticipants believed that Bonn is the largest German city; it is 23rd.) Thus,the demonstration of a less-is-more effect is particularly difficult in thissituation because the recognition heuristic only makes predictions aboutuncertain inference, not about the kinds of definite knowledge the Amer-icans had.

Results

The American participants scored a mean 71.1% (median71.0%) correct on their own cities. On the German cities, the meanaccuracy was 71.4% (median 73.0%) correct. Despite the presence ofdefinite knowledge about the American cities, the recognition heuris-tic still caused a slight less-is-more effect. For half of the participants,we kept track of which cities they recognized, as in previous exper-iments. For this group, the mean proportion of inferences in accor-dance with the recognition heuristic was 88.5% (median 90.5%). Therecognition test showed that participants recognized a mean of 12German cities, roughly half of the total, which indicates that they wereable to apply the recognition heuristic nearly as often as theoreticallypossible (see Equation 1).

In a study that is somewhat the reverse of this one, a less-is-more effect was demonstrated with German students who scoredhigher when tested on American cities than on German ones(Hoffrage, 1995; see also Gigerenzer, 1993).

Despite all the knowledge—including certain knowledge—theAmericans had about their own cities, and despite their limitedknowledge about Germany, they could not make more accurateinferences about American cities than about German ones. Facedwith German cities, the participants could apply the recognitionheuristic. Faced with American cities, they had to rely on knowl-edge beyond recognition. The fast and frugal recognition heuristicexploited the information inherent in a lack of knowledge to makeinferences that were slightly more accurate than those achievedfrom more complete knowledge.

Will a Less-Is-More Effect Occur as RecognitionKnowledge Is Acquired?

“A little learning is a dangerous thing,” warned Alexander Pope.The recognition heuristic predicts cases in which increases inknowledge can lead to decreases in inferential accuracy. Equa-tion 1 predicts that if � � �, the proportion of accurate inferenceswill increase up to a certain point when a person’s knowledgeincreases but thereafter decrease because of the diminishing ap-plicability of the recognition heuristic. This study aims to demon-strate that less-is-more effects can emerge over the course of timeas ignorance is replaced with recognition knowledge.

The design of the experiment was as follows. German partici-pants came to the laboratory four times and were tested on Amer-ican cities. As they were tested repeatedly, they may have gainedwhat we lightheartedly call an “experimentally induced” sense ofrecognition for the names of cities they had not recognized beforethe experiment. This induced recognition is similar to that gener-ated in the “overnight fame” experiments by Jacoby, Kelley,Brown, and Jasechko (1989), in which mere exposure causednonfamous names to be judged as famous. Can mere exposure tocity names cause people to infer that formerly unrecognized citiesare large? If so, this should cause accuracy on certain questions todrop. For instance, in the first session, a German who has heard ofDallas but not Indianapolis, would correctly infer that Dallas islarger. However, over the course of repeated testing, this personmay develop an experimentally induced sense of recognition forIndianapolis without realizing it. Recognizing both cities, shebecomes unable to use the recognition heuristic and may have toguess.

It is difficult to produce the counterintuitive effect that accuracywill decrease as city names are learned because it is contingent onseveral assumptions. The first is that recognition will be experi-mentally induced, and there is evidence for this phenomenon in thework by Jacoby, Kelley, et al. (1989) and Jacoby, Woloshyn, andKelley (1989). The second assumption is that people use therecognition heuristic, and there is evidence for this in the experi-mental work reported here. The third assumption is that, for theseparticipants, the recognition validity � is larger than the knowledgevalidity �: a necessary condition for a less-is-more effect. Thesimulation depicted in Figure 3 indicates that this condition mighthold for certain people.

Method

Participants were 16 residents of Munich, Germany, who were paid fortheir participation. (They were not paid, however, for the correctness oftheir answers because this would have encouraged them to do research on


populations between sessions). In the first session, after a practice test toget used to the computer, they were shown the names of the 75 largestAmerican cities in random order. (The first three, New York, Los Angeles,and Chicago, were excluded because many Germans know they are thethree largest). For each city, participants indicated whether they had heardof the city before the experiment, and then, to encourage the encoding ofthe city name in memory, they were asked to write the name of each cityas it would be spelled in German. Participants were not informed that thecities were among the largest in the country, only that they were Americancities. They were then given a test consisting of 300 pairs of cities,randomly drawn for each participant, and asked to choose the larger city ineach pair.

About 1 week later, participants returned to the lab and took another testof 300 pairs of American cities randomly drawn for each participant. Thethird week was a repetition of the second week. The fourth week was thecritical test. This time, participants were given 200 carefully selectedquestions. There were two sets of 100 questions each that were used to testtwo predictions. The first set of 100 was composed of questions taken fromthe first week’s test. This set was generated by listing all questions from thefirst session’s test in which one city was recognized and the other not(according to each participant’s recognition in the first week) and randomlydrawing (with replacement) 100 times. We looked at these repeated ques-tions to test the prediction that accuracy will decrease as recognitionknowledge is acquired.

A second set of 100 questions consisted of pairs with one “experimen-tally induced” city (i.e., a city that was unrecognized before the first sessionbut may have become recognized over the course of repeated testing) anda new, unrecognized city introduced for the first time in the fourth session.All new cities were drawn from the next 50 largest cities in the UnitedStates, and a posttest recognition survey was used to verify whether theywere novel to the participants. (Participants, however, did not know fromwhich source any of the cities in the experiment were drawn.)

Which would people choose: an experimentally induced city that theyhad learned to recognize in the experiment or a city they had never heardof before? If people use the recognition heuristic, this choice should not berandom but should show a systematic preference for the experimentallyinduced cities. This set of questions was introduced to test the hypothesisthat recognition information acquired during the experiment would be usedas a surrogate for genuine recognition information.

Results

If the assumptions just specified hold, one should observe thatthe percentage of times experimentally induced cities are inferredto be larger than recognized cities should increase and, subse-quently, cause accuracy to decrease.

In the first week, participants chose unrecognized cities overrecognized ones in 9.6% of all applicable cases, consistent with theproportions reported in Studies 1 and 3. By the fourth week, theexperimentally induced cities were chosen over those that wererecognized before the first session 17.2% of the time (Table 1).Participants’ accuracy on the 100 repeated questions dropped froma mean of 74.8% correct (median 76%) in the first week to 71.3%correct (median 74%) in the fourth week, t(15) � 1.82, p � .04,one-tailed. To summarize, as unrecognized city names were pre-sented over 4 weeks, participants became more likely to infer thatthese cities were larger than recognized cities. As a consequence,accuracy dropped during this month-long experiment. Surpris-ingly, this occurred despite the participants having ample time tothink about American cities and their populations, to recall infor-mation from memory, to ask friends or look in reference books forcorrect answers, or to notice stories in the media that could informtheir inferences.

How did participants make inferences in the second set of 100questions, in which each question consisted of two unrecognizedcities: one new to the fourth session and one experimentallyinduced? If the repeated presentation of city names had no effecton inferences, participants would be expected to choose both typesof cities equally often (about 50% of the time) in the fourthsession. However, in the fourth session, participants chose theexperimentally induced cities over the new ones 74.3% of the time(median 77%). Recognition induced in the laboratory had amarked effect on the direction of people’s inferences.

The Ecological Rationality of Name Recognition

What is the origin of the recognition heuristic as a strategy? Inthe case of avoiding food poisoning, organisms seem as if they aregenetically prepared to act in accordance with the recognitionheuristic. Wild Norway rats do not need to be taught to preferrecognized foods over novel ones; food neophobia is instinctual(Barnett, 1963). Having such a strategy as an instinct makesadaptive sense: If an event is life threatening, organisms needing tolearn to follow the recognition heuristic will most likely die beforethey get the chance. Learning the correlation between name rec-ognition and city size is not an adaptive task. How did the asso-ciation between recognition and city population develop in theminds of the participants we investigated?

The recognition validity—the strength of the association be-tween recognition and the criterion—can be explained as a func-tion of the ecological and the surrogate correlations that connect anunknown environment with the mind by means of a mediator (seeFigure 1). If the media are responsible for the set of proper nameswe recognize, the number of times a city is mentioned in thenewspapers should correlate with the proportion of readers whorecognize the city; that is, there should be a substantial surrogatecorrelation. Furthermore, there should be a strong ecological cor-relation (larger cities should be mentioned in the news more often).To test this postulated ecological structure, we analyzed two news-papers with large readerships: the Chicago Tribune in the UnitedStates and Die Zeit in Germany.

Method

Using an Internet search tool, we counted the number of articles pub-lished in the Chicago Tribune between January 1985 and July 1997 inwhich the words Berlin and Germany were mentioned together. Therewere 3,484. We did the same for all cities in Germany with more than

Table 1A Less-Is-More Effect Resulting From Learning

SessionMean %correct

Median %correct Inferences (%)a

1 74.8 76 9.64 71.3 74 17.2

Note. The percentage of correct answers drops from the first to the fourthsessions. As German participants saw the same novel American city namesover and over again in repeated testing, they began to choose them overcities that they recognized from the real world (column 4).a Represents inferences in which an experimentally induced city was cho-sen over one recognized from before the experiment.


100,000 inhabitants (there are 83 of them) and checked under manypossible spellings and misspellings. Sixty-seven Chicago residents weregiven a recognition test in which they indicated whether they had heard ofeach of the German cities. The proportion of participants who recognizeda given city was the city’s recognition rate.

The same analysis was performed with a major German languagenewspaper, Die Zeit. For each American city with more than 100,000inhabitants, the number of articles was counted in which it was mentioned.The analysis covered the period from May 1995 to July 1997. Recognitiontests for the American cities were administered to 30 University of Salz-burg students (Hoffrage, 1995).

Results

Three measures were obtained for each city: the actual popula-tion, the number of mentions in the newspaper, and the recognitionrate. Figure 7 illustrates the underlying structure and shows thecorrelations between these three measures. For the class of allGerman cities with more than 100,000 inhabitants, the ecologicalcorrelation (i.e., the correlation between the number of newspaperarticles in the Chicago Tribune mentioning a city and its actualpopulation) was .70; the number of times a city is mentioned in thenewspaper is a good indicator of the city’s population. The surro-gate correlation, that is, the correlation between the number ofnewspaper articles about a city and the number of people recog-nizing it was .79; the recognition rates are more closely associatedwith the number of newspaper articles than with the actual popu-lations. This effect is illustrated by large German and Americancities that receive little newspaper coverage, such as Essen, Dort-mund, and San Jose. Their recognition rates tend to follow the lowfrequency of newspaper citations rather than their actual popula-tion. Finally, the correlation between the number of people recog-nizing a city and its population—the recognition validity expressedas a correlation—was .60. Recognition is a good predictor ofactual population but not as strong as the ecological and surrogatecorrelations.

Do these results stand up in a different culture? For the class ofall American cities with more than 100,000 inhabitants, the eco-logical correlation was .72. The surrogate correlation was .86, andthe correlation between recognition and the rank order of citieswas .66. These results are consistent with those from the Americandata, with slightly higher correlations.

This study illustrates how to analyze the ecological rationality ofthe recognition heuristic. The magnitude of the recognition valid-ity, together with n and N (Equation 1), specifies the expectedaccuracy of the heuristic but does not explain why recognition isinformative. The ecological and surrogate correlations (see Figure1) allow one to model the network of mediators that could explainwhy and when a lack of recognition is informative, that is, whenmissing knowledge is systematically rather than randomlydistributed.

Institutions, Firms, and Name Recognition

For both the Chicago Tribune and Die Zeit, the surrogate cor-relation was the strongest association, which suggests that individ-ual recognition was more in tune with the media than with theactual environment. Because of this, to improve the perceivedquality of a product, a firm may opt to manipulate the product’sname recognition through advertising instead of investing in theresearch and development necessary to improve the product’squality. Advertising manipulates recognition rates directly and isone way in which institutions exploit recognition-based inference.

One way advertisers achieve name recognition is by associatingthe name of their products with strong visual images. If the realpurpose of the ad is to convey name recognition and not commu-nicate product information, then only the attention-getting qualityof the images, and not the content of the images, should matter.Advertiser Oliviero Toscani bet his career on this strategy. In hiscampaign for Bennetton, he produced a series of advertisementsthat conveyed nothing about actual Bennetton products but soughtto induce name recognition by association with shocking images,such as a corpse lying in a pool of blood or a dying AIDS patient.Would associating the name of a clothing manufacturer withbloody images cause people to learn the name? Toscani (1997)reported that the campaign was a smashing success and that itvaulted Bennetton’s name recognition into the top five in theworld, even above that of Chanel. In social domains, recognition isoften correlated with quality, resources, wealth, and power. Why?Perhaps because individuals with these characteristics tend to bethe subject of news reporting as well as gossip. As a result,advertisers pay great sums for a place in the recognition memoryof the general public, and the lesser known people, organizations,institutions, and nations of the world go on crusades for namerecognition. They all operate on the principle that if we do notrecognize them, then we will not favor them.

Those who try to become recognized through advertising maynot be wasting their resources. As mentioned, the “overnightfame” experiments by Jacoby, Kelley, et al. (1989) demonstratethat people may have trouble distinguishing between names theylearned to recognize in the laboratory and those they learned in thereal world. In a series of classic studies, participants read famousand nonfamous names, waited overnight, and then made famejudgments. Occasionally, people would mistakenly infer that anonfamous name, learned in the laboratory, was the name of a

Figure 7. Ecological correlation, surrogate correlation, and recognitioncorrelation. The first value is for American cities and the German news-paper Die Zeit as mediator, and the second value is for German cities andthe Chicago Tribune as mediator. Note that the recognition validity isexpressed, for comparability, as a correlation (between the number ofpeople who recognize the name of a city and its population).


famous person. The recognition heuristic can be fooled, and thepreviously anonymous can inherit the appearance of being skilled,famous, important, or powerful.

Can Compensatory Strategies Mimicthe Recognition Heuristic?

So far we have dealt only with recognition and not with infor-mation recalled from memory. The recognition heuristic can act asa subroutine in heuristics that process knowledge beyond recog-nition. Examples are the Take The Last and Take The Bestheuristics, which feature the recognition heuristic as the first step(Gigerenzer & Goldstein, 1996). These heuristics are also non-compensatory and can benefit from the information implicit in alack of recognition but are also able to function with completeinformation. However, in cases in which one object is recognizedand the other is not, they make the same inference as the recog-nition heuristic.

Do compensatory strategies exist that do not have the recogni-tion heuristic as an explicit step but that, unwittingly, make infer-ences as if they did? It seems that only noncompensatory cognitivestrategies (such as lexicographic models) could embody the non-compensatory recognition heuristic. However, it turns out that thisis not so. For instance, consider a simple tallying strategy thatcounts the pieces of positive evidence for each of two objects andchooses the object with more positive evidence (see Gigerenzer &Goldstein, 1996). Such a strategy would be a relative of a strategythat searches for confirming evidence and ignores disconfirmingevidence (e.g., Klayman & Ha, 1987). Interestingly, because therecan be no positive evidence for an unrecognized object, such astrategy will never choose an unrecognized object over a recog-nized one. If this strategy, however, were, in addition, to payattention to negative evidence and choose the object with the largersum of positive and negative values, it would no longer make thesame choice as the recognition heuristic. This can be seen from thefact that a recognized object may carry more negative than positiveevidence, whereas an unrecognized object carries neither. As aconsequence, compensatory strategies such as tallying that mimicthe recognition heuristic will also mimic a less-is-more effect(Gigerenzer & Goldstein, 1999).

Domain Specificity

The recognition heuristic is not a general purpose strategy forreasoning. It is a domain-specific heuristic. Formally, its domainspecificity can be defined by two characteristics. First, someobjects must be unrecognized (n � N). Second, the recognitionvalidity must be higher than chance (� � .5). However, whatdomains, in terms of content, have these characteristics? We beginwith examples of recognition validities � � .5, and then proposeseveral candidate domains, without claims to an exhaustive list.

Table 2 lists 10 topics of general knowledge, ranging from theworld’s highest mountains to the largest American banks. Wouldthe recognition heuristic help to infer correctly which of, forexample, two mountains is higher? We surveyed 20 residents ofBerlin, Germany, about which objects they recognized for each ofthe 10 topics and then computed each participant’s recognitionvalidity for each topic. Table 2 shows the average recognition

validities. These validities are often high, and the topics in Table 2can be extended to include many others.

Domains

Alliances and competition. Deciding whom to befriend andwhom to compete against is an adaptive problem for humans andother social animals (e.g., Cosmides & Tooby, 1992; de Waal,1982). Organisms need a way to assess quickly who or what in theworld has influence and resources. Recognition of a person, group,firm, institution, city, or nation signals its social, political, andeconomic importance to others of its kind that may be contem-plating a possible alliance or competition. For instance, in thehighly competitive stock market, the recognition heuristic has, onaverage, matched or outperformed major mutual funds, the market,randomly picked stocks, and the less recognized stocks. This resulthas been obtained both for the American and the German stockmarkets (Borges, Goldstein, Ortmann, & Gigerenzer, 1999).

Risk avoidance. A second domain concerns risky behaviorwhen the risks associated with objects are too dangerous to belearned by experience or too rare to be learned in a lifetime(Cosmides & Tooby, 1994). Food avoidance is an example; therecognition heuristic helps organisms avoid toxic foods (Galef,1987). If an organism had to learn from individual experiencewhich risks to take (or foods to eat), it could be fatal. In many wildenvironments, novelty carries risk, and recognition can be a simpleguide to a safe choice.

Social bonding. Social species (i.e., species that have evolvedcognitive adaptations such as imprinting on parents by their chil-dren, coalition forming, and reciprocal altruism) are equipped witha powerful perceptual machinery for the recognition of individualconspecifics. Face recognition and voice recognition are examples.This machinery is the basis for the use of the recognition heuristicas a guide to social choices. For instance, one author’s daughteralways preferred a babysitter whom she recognized to one ofwhom she had never heard, even if she was not enthusiastic aboutthe familiar babysitter. Recognition alone cannot tell us whom,among recognized individuals, to trust; however, it can suggestthat we not trust unrecognized individuals.

Fast and Frugal Heuristics

The current work on the recognition heuristic is part of aresearch program to study the architecture and performance of fast

Table 2A Sample of Recognition Validities

Topic Recognition validity �

10 Largest Indian cities 0.9520 Largest French cities 0.8710 Highest mountains 0.8515 Largest Italian provinces 0.8110 Largest deserts 0.8010 Tallest buildings 0.7910 Largest islands 0.7910 Longest rivers 0.6910 Largest U.S. banks 0.6810 Largest seas 0.64


and frugal heuristics (Gigerenzer & Goldstein, 1996; Gigerenzer &Selten, 2001; Gigerenzer, Todd, & the ABC Research Group,1999; Goldstein & Gigerenzer, 1999; Todd & Gigerenzer, 2000).The recognition heuristic is a prototype of these adaptive tools. Ituses recognition, a capacity that evolution has shaped over mil-lions of years that allows organisms to benefit from their ownignorance. The heuristic works with limited knowledge and limitedtime and even requires a certain amount of missing information.Other research programs have pointed out that one’s ignorance canbe leveraged to make inferences (Glucksberg & McCloskey,1981); however, they refer to an ignorance of deeper knowledge,not ignorance in the sense of a mere lack of recognition. Theprogram of studying fast and frugal heuristics is based on thefollowing theoretical principles.

Rules for Search, Stopping, and Decision

The common defining characteristics of these heuristics are fastand frugal rules for (a) search, that is, where to search for cues; (b)stopping, that is, when to stop searching without attempting tocompute an optimal stopping point at which the costs of furthersearch exceed the benefits; and (c) decision, that is, how to makean inference or decision after search is stopped. The recognitionheuristic follows particularly simple rules. Search extends only torecognition information, not to recall. Search is stopped wheneverone object is recognized and the other is not; no further informa-tion is looked up about the recognized object. The simple decisionrule is to choose the recognized object. Limited search and non-optimizing stopping rules are the key processes in Simon’s (e.g.,1955) and Selten’s (e.g., 1998) models of bounded rationality.However, only a few theories of cognitive processing specifymodels of search and stopping; this even holds for those whoattach the label “bounded rationality” to their ideas. Most theoriesonly specify rules for decision, such as how information should becombined in some linear or nonlinear fashion.

Transparency

Fast and frugal heuristics are formulated in a way that is highlytransparent: It is easy to discern and understand just how theyfunction in making decisions. Because they involve few, if any,free parameters and a minimum of computation, each step of thealgorithm is open to scrutiny. These simple heuristics stand insharp contrast to more complex and computationally involvedmodels of mental processes that may generate good approxima-tions to human behavior but are also rather opaque. For instance,the resurgence of connectionism in the 1980s brought forth a cropof neural networks that were respectable models for a variety ofpsychological phenomena but whose inner workings remainedmysterious even to their creators (for alternatives see Regier, 1996;Rumelhart & Todd, 1993). Whereas these models are preciselyformulated but nontransparent, there also exist imprecisely formu-lated and nontransparent models of heuristics, such as “represen-tativeness,” “availability,” and other one-word explanations (seeGigerenzer, 1996, 1998; Lopes, 1991). Transparent fast and frugalheuristics use no (or only a few) free parameters, allow one tomake quantifiable and testable predictions, and avoid possiblemisunderstanding (or mystification) of the processes involved,even if they do sacrifice some of the allure of the unknown.

Ecological Rationality

Models of cognitive algorithms—from models of risky choice(e.g., Lopes, 1994) to multiple regression lens models (e.g., Ham-mond, Hursch, & Todd, 1964) to neural networks (e.g., Regier,1996)—typically rely on predictor variables that have objectivecorrespondents in the outside world. In this article, we proposedand studied a heuristic that makes inferences without relying onobjective predictors but solely on an internal lack of recognition.

We defined the conditions in which the recognition heuristic isapplicable and in which the less-is-more effect emerges. In theexperimental studies reported, inferences accorded with the rec-ognition heuristic about 90% of the time or more. The recognitionheuristic, in the form of Equation 1, predicts how changes in testsize affect accuracy, and these predicted changes were experimen-tally confirmed in 26 of 28 cases. The heuristic also predictsnoncompensatory judgments, and this prediction was obtained in atraining experiment in which participants were taught predictiveinformation that suggested contradicting recognition. We demon-strated two less-is-more effects. In the first, a group achievedslightly higher accuracy in a foreign domain in which they knewlittle than in a familiar domain where they knew far more. In thesecond, we saw how experimentally induced recognition signifi-cantly affects the inferences people make.

The recognition heuristic is a cognitive adaptation. In cases ofextremely limited knowledge, it is perhaps the only strategy anorganism can follow. However, it is also adaptive in the sense thatthere are situations, called less-is-more effects, in which the rec-ognition heuristic results in more accurate inferences than a con-siderable amount of knowledge can achieve. It is the model of anecologically rational heuristic, exploiting patterns of informationin the environment to make accurate inferences in a fast and frugalway.

References

Ayton, P., & Onkal, D. (1997). Forecasting football fixtures: Confidenceand judged proportion correct. Unpublished manuscript, Department ofPsychology, The City University of London.

Barnett, S. A. (1963). The rat: A study in behavior. Chicago: Aldine.Birnbaum, M. H. (1983). Base rates in Bayesian inference: Signal detection

analysis of the cab problem. American Journal of Psychology, 96,85–94.

Borges, B., Goldstein, D. G., Ortmann, A., & Gigerenzer, G. (1999). Canignorance beat the stock market? In G. Gigerenzer, P. M. Todd, & theABC Research Group (Eds.), Simple heuristics that make us smart (pp.59–72). New York: Oxford University Press.

Brown, J., Lewis, V. J., & Monk, A. F. (1977). Memorability, wordfrequency and negative recognition. Quarterly Journal of ExperimentalPsychology, 29, 461–473.

Cosmides, L., & Tooby, J. (1992). Cognitive adaptations for social ex-change. In J. Barkow, L. Cosmides, & J. Tooby (Eds.), The adaptedmind: Evolutionary psychology and the generation of culture (pp. 163–228). New York: Oxford University Press.

Cosmides, L., & Tooby, J. (1994). Beyond intuition and instinct blindness:Toward an evolutionarily rigorous cognitive science. Cognition, 50,41–77.

Craik, F. I. M., & McDowd, M. (1987). Age differences in recall andrecognition. Journal of Experimental Psychology: Learning, Memory,and Cognition, 14, 474–479.

Delbecq-Derousne, J., Beauvois, J. F., & Shallice, T. (1990). Preservedrecall versus impaired recognition. Brain, 113, 1045–1074.


de Waal, F. (1982). Chimpanzee politics. London: Jonathan Cape.Duncker, K. (1945). On problem solving (L. S. Lees, Trans.). Psycholog-

ical Monographs, 58 (5, Whole No. 270). (Original work published1935)

Edwards, W. (1968). Conservatism in human information processing. In B.Kleinmuntz (Ed.), Formal representation of human judgment (pp. 17–52). New York: Wiley.

Einhorn, H. J. (1970). The use of nonlinear, noncompensatory models indecision making. Psychological Bulletin, 73, 221–230.

Einhorn, H. J., & Hogarth, R. M. (1981). Behavioral decision theory:Processes of judgment and choice. Annual Review of Psychology, 32,53–88.

Fishburn, P. C. (1974). Lexicographic orders, utilities and decision rules: Asurvey. Management Science, 20, 1442–1471.

Galef, B. G. (1987). Social influences on the identification of toxic foodsby Norway rats. Animal Learning & Behavior, 15, 327–332.

Galef, B. G., McQuoid, L. M., & Whiskin, E. E. (1990). Further evidencethat Norway rats do not socially transmit learned aversions to toxic baits.Animal Learning Behavior, 18, 199–205.

Gigerenzer, G. (1991). From tools to theories: A heuristic of discovery incognitive psychology. Psychological Review, 98, 254–267.

Gigerenzer, G. (1993). The bounded rationality of probabilistic mentalmodels. In K. I. Manktelow & D. E. Over (Eds.), Rationality: Psycho-logical and philosophical perspectives (pp. 284–313), London: Rout-ledge.

Gigerenzer, G. (1996). On narrow norms and vague heuristics: A reply toKahneman and Tversky. Psychological Review, 103, 592–596.

Gigerenzer, G. (1998). Surrogates for theories. Theory & Psychology, 8,195–204.

Gigerenzer, G. (2000). Adaptive thinking: Rationality in the real world.New York: Oxford University Press

Gigerenzer, G., & Goldstein, D. G. (1996). Reasoning the fast and frugalway: Models of bounded rationality. Psychological Review, 104, 650–669.

Gigerenzer, G., & Goldstein, D. G. (1999). Betting on one good reason:The take the best heuristic. In G. Gigerenzer, P. M. Todd, & the ABCResearch Group (Eds.), Simple heuristics that make us smart (pp. 75–95). New York: Oxford University Press.

Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reason-ing without instruction. Frequency formats. Psychological Review, 102,684–704.

Gigerenzer, G., & Murray, D. J. (1987). Cognition as intuitive statistics.Hillsdale, NJ: Erlbaum.

Gigerenzer, G., & Selten, R. (Eds.). (2001). Bounded rationality: Theadaptive toolbox. Cambridge, MA: MIT Press.

Gigerenzer, G., Todd, P. M., & the ABC Research Group. (1999). Simpleheuristics that make us smart. New York: Oxford University Press.

Glucksberg, S., & McCloskey, M. (1981). Decisions about ignorance:Knowing that you don’t know. Journal of Experimental Psychology:Human Learning and Memory, 7, 311–325.

Goldstein, D. G. (1997). Models of bounded rationality for inference.Dissertation Abstracts International, 58 (01), 435B. (UMI No. AAT9720040)

Goldstein, D. G., & Gigerenzer, G. (1999). The recognition heuristic: Howignorance makes us smart. In G., Gigerenzer, P. M. Todd, & the ABCResearch Group (Eds.), Simple heuristics that make us smart (pp. 37–58). New York: Oxford University Press.

Griggs, R. A., & Cox, J. R. (1982). The elusive thematic-materials effectin Wason’s selection task. British Journal of Psychology, 73, 407–420.

Groner, M., Groner, R., & Bischof, W. F. (1983). Approaches to heuristics:A historical review. In R. Groner, M. Groner, & W. F. Bischof (Eds.),Methods of heuristics (pp. 1–18). Hillsdale, NJ: Erlbaum.

Hammond, K. R. (1996). Human judgment and social policy. Irreducible

uncertainty, inevitable error, unavoidable injustice. New York: OxfordUniversity Press.

Hammond, K. R., Hursch, C. J., & Todd, F. J. (1964). Analyzing thecomponents of clinical inference. Psychological Review, 71, 438–456.

Hoffrage, U. (1995). The adequacy of subjective confidence judgments:Studies concerning the theory of probabilistic mental models. Unpub-lished doctoral dissertation, University of Salzburg, Austria.

Hogarth, R. M. (1987). Judgement and choice: The psychology of decision.Chichester, United Kingdom: Wiley.

Jacoby, L. L., Kelley, C., Brown, J., & Jasechko, J. (1989). Becomingfamous overnight: Limits on the ability to avoid unconscious influencesof the past. Journal of Personality and Social Psychology, 56, 326–338.

Jacoby, L. J., Woloshyn, V., & Kelley, C. (1989). Becoming famouswithout being recognized: Unconscious influences of memory producedby dividing attention. Journal of Experimental Psychology, 118, 115–125.

Kahneman, D., & Tversky, A. (1996). On the reality of cognitive illusions.Psychological Review, 103, 582–591.

Keeney, R. L., & Raiffa, H. (1993). Decisions with multiple objectives.Cambridge, United Kingdom: Cambridge University Press.

Kelley, H. H. (1973). The process of causal attribution. American Psychol-ogist, 28, 107–128.

Klayman, J., & Ha, Y. (1987). Confirmation, disconfirmation, and infor-mation in hypothesis testing. Psychological Review, 94, 211–228.

Lopes, L. L. (1991). The rhetoric of irrationality. Theory and Psychol-ogy, 1, 65–82.

Lopes, L. L. (1994). Psychology and economics: Perspectives on risk,cooperation, and the marketplace. Annual Review of Psychology, 45,197–227.

McKenna, P., & Warrington, E. K. (1980). Testing for nominal dysphasia.Journal of Neurology, Neurosurgery and Psychiatry, 43, 781–788.

Mellers, B. A., Schwartz, A., & Cooke, A. D. J. (1998). Judgment anddecision making. Annual Review of Psychology, 49, 447–477.

Payne, J. W., Bettman, J. R., & Johnson, E. J. (1993). The adaptivedecision maker. New York: Cambridge University Press.

Piattelli-Palmarini, M. (1994). Inevitable illusions. How mistakes of reasonrule our minds. New York: Wiley.

Polya, G. (1954). Mathematics and plausible reasoning. Vol. 1: Inductionand analogy in mathematics. Princeton, NJ: Princeton University Press.

Regier, T. (1996). The human semantic potential: Spatial language andconstrained connectionism. Cambridge, MA: MIT Press.

Rumelhart, D. E., & Todd, P. M. (1993). Learning and connectionistrepresentations. In D. E. Meyer & S. Kornblum (Eds.), Attention andperformance XIV (pp. 3–30). Cambridge, MA: MIT Press/BradfordBooks.

Schacter, D. L., & Tulving, E. (1994). What are the memory systems of1994? In D. L. Schacter & E. Tulving (Eds.), Memory systems (pp.1–38). Cambridge, MA: MIT Press.

Schonfield, D., & Robertson, B. (1966). Memory storage and aging.Canadian Journal of Psychology, 20, 228–236.

Selten, R. (1998). Aspiration adaptation theory. Journal of MathematicalPsychology, 42, 191–214.

Semenza, C., & Sgaramella, T. M. (1993). Production of proper names: Aclinical case study of the effects of phonemic cueing. In G. Cohen &D. M. Burke (Eds.), Memory for proper names (pp. 265–280). Hove,United Kingdom: Erlbaum.

Semenza, C., & Zettin, M. (1989). Evidence from aphasia for the role ofproper names as pure referring expressions. Nature, 342, 678–679.

Shepard, R. N. (1967a). On subjectively optimum selections among multi-attribute alternatives. In W. Edwards & A. Tversky (Eds.), Decisionmaking (pp. 257–283) Harmondsworth: Penguin Books.

Shepard, R. N. (1967b). Recognition memory for words, sentences, andpictures. Journal of Verbal Learning and Verbal Behavior, 6, 156 –163.


Simon, H. A. (1955). A behavioral model of rational choice. QuarterlyJournal of Economics, 69, 99–118.

Squire, L. R., Knowlton, B., & Musen, G. (1993). The structure andorganization of memory. Annual Review of Psychology, 44, 453–495.

Standing, L. (1973). Learning 10,000 pictures. Quarterly Journal of Ex-perimental Psychology, 25, 207–222.

Strack, F., & Bless, H. (1994). Memory for nonoccurrences: Metacognitiveand presuppositional strategies. Journal of Memory and Language, 33,203–217.

Todd, P. N., & Gigerenzer, G. (2000). Precis of simple heuristics that makeus smart. Behavioral and Brain Sciences, 23, 727–780.

Toscani, O. (1997). Die Werbung ist ein lachelndes Aas [Advertisement isa beast that smiles at us]. Frankfurt, Germany: Fischer.

Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heu-ristics and biases. Science, 185, 1124–1131.

Warrington, E. K., & McCarthy, R. A. (1988). The fractionation of retro-grade amnesia. Brain and Cognition, 7, 184–200.

Received December 15, 1998Revision received September 20, 2000

Accepted June 12, 2001 �


models of ecological rationality: the recognition · pdf file · 2003-12-31models...

Documents